Back to Models

MiniMax M2.5

minimaxminimax/m2.5

MiniMax M2.5 is an agent-native frontier model trained to reason efficiently, decompose tasks optimally, and complete complex workflows under real-world constraints. Combines high inference throughput with RL-focused token-efficient reasoning. Suited for full-stack software projects, research workflows, long-horizon planning, and multi-tool orchestration.

MiniMax M2.5 is an open-weight frontier language model released on February 12, 2026, as the third iteration of the M2 family (M2 -> M2.1 -> M2.5) in just three and a half months. Built on a 230B-parameter Mixture-of-Experts (MoE) architecture with 10B activated parameters per token, M2.5 is trained via a novel agent-native RL framework called Forge using hundreds of thousands of simulated real-world environments. It achieves 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench (first place at release), and 76.3% on BrowseComp -- placing it at the frontier alongside Claude Opus 4.6, at roughly 10-20x lower cost.

For a TheRouter operator, M2.5 is the highest-value coding and agentic model on the market per dollar. It completes SWE-Bench Verified tasks 37% faster than M2.1, matching Opus 4.6's end-to-end speed, while total cost per task is only 10% that of Opus 4.6. M2.5 ships with a Modified MIT license allowing commercial use with attribution, and is available in two variants: standard (50 TPS) and Lightning (100 TPS). Its successor M2.7 (April 2026) has since shipped; for brand-new agent projects evaluate M2.7 first, but M2.5 remains the most battle-tested and thoroughly documented release in the family, and the one that broke the cost barrier for frontier-level agentic coding.

Best for
  • β€’ Agentic coding and software engineering -- M2.5 achieves 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench, edging out Claude Opus 4.6 on both Droid (79.7% vs 78.9%) and OpenCode (76.1% vs 75.9%) agent harnesses
  • β€’ Full-stack cross-platform development across Web, Android, iOS, and Windows -- handles everything from 0-to-1 system design through feature iteration to code review and system testing
  • β€’ Multi-language programming -- trained on Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby across 200,000+ real-world environments
  • β€’ Cost-sensitive high-throughput agent operations -- at $0.30/hour (50 TPS) continuous uptime, suitable for always-on autonomous agents
  • β€’ Office and finance productivity -- Word document creation, PowerPoint design, Excel financial modeling, achieving 59.0% win rate on GDPval-MM against other mainstream models
Reach for something else if
  • β€’ Multimodal tasks -- M2.5 is text-only; no native image input, vision, or audio modalities. Pair with a vision model for screenshots or diagrams
  • β€’ Ultra-long generation beyond 64K tokens -- max_completion is 65,536 tokens; for uninterrupted document generation beyond that, plan chunking or use a higher-max-output model
  • β€’ Scenarios with existing workflow tuned on M2.1 behavior -- M2.5 has an emergent 'architect mindset' (pre-planning structure before writing code) that can change agent behavior. Validate on benchmarks before migrating production workloads
Context Length
205K
Max Output
66K
Input Priceper 1M tokens
$0.324/ 1M tokens
Output Priceper 1M tokens
$1.30/ 1M tokens

Modalities

text→text

Pricing Breakdown

TypeRate
Input$0.324 / 1M tokens
Output$1.30 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Release dateFebruary 12, 2026www.minimax.io β†—verified
ArchitectureMixture of Experts (MoE) -- 230B total / 10B activated per tokenwww.minimax.io β†—verified
LicenseModified MIT -- commercial use allowed with attribution (display 'MiniMax M2.5' on UI)github.com β†—verified
RL frameworkForge -- agent-native RL framework with CISPO algorithmwww.minimax.io β†—verified
Training environments200,000+ real-world simulated environmentswww.minimax.io β†—verified
Training languagesGo, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Rubywww.minimax.io β†—verified

Benchmarks

BenchmarkDistributionScoreSource
SWE-bench Verified
Tested with Claude Code scaffolding, averaged over 4 runs. Also achieves 79.7% on Droid scaffold (vs Opus 4.6 78.9%) and 76.1% on OpenCode scaffold (vs Opus 4.6 75.9%)
80.2%www.minimax.io β†—
Multi-SWE-bench
First place at release -- multi-language software engineering benchmark
51.3%www.minimax.io β†—
BrowseComp
With context management. Tests browsing and autonomous web search capabilities
76.3%www.minimax.io β†—
BFCL (Tool Calling)
Berkeley Function Calling Leaderboard -- high-precision agentic workflows
76.8%www.minimax.io β†—
VIBE-Pro
Internal MiniMax benchmark. Upgraded to Pro version with M2.5. Performance on par with Opus 4.5
SOTA-levelwww.minimax.io β†—
SWE-bench Multilingual
Multi-language coding capability across 10+ languages
72.5%www.minimax.io β†—
Wide Search
Multi-step information retrieval. Tested with WebExplorer agent framework
Industry-leadingwww.minimax.io β†—
RISE
Realistic Interactive Search Evaluation -- internal MiniMax benchmark for expert-level multi-step search on information-dense webpages
Expert-levelwww.minimax.io β†—
GDPval-MM
Internal Cowork Agent evaluation -- pairwise comparisons against other mainstream models
59.0% avg win ratewww.minimax.io β†—

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "minimax/m2.5",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

Chat completion

Standard OpenAI-compatible chat completion. Supports tool calling, JSON mode, streaming, and stop sequences.

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/m2.5",
    "messages": [
      {"role": "system", "content": "You are an expert software engineer."},
      {"role": "user", "content": "Write a Python function to merge two sorted lists."}
    ],
    "temperature": 0.3,
    "max_tokens": 2048
  }'

More from minimax

Similar models

Cross-provider sibling models

News & changes

2026-02-12

MiniMax releases M2.5 -- a frontier model matching Claude Opus 4.6 at 1/10th the cost

MiniMax M2.5 achieves 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench (first place at release), with BrowseComp at 76.3%. Trained via the Forge agent-native RL framework on 200,000+ simulated environments. The model is open-weight (Modified MIT) and priced at $0.15/$1.15 per million input/output tokens via MiniMax API, or through TheRouter at $0.40/$1.60 per million tokens.

re-authored by TheRouterwww.minimax.io β†—

Frequently asked

How does M2.5 compare to M2.1 and M2.7?

M2.5 is a significant upgrade over M2.1: +6.2% on SWE-Bench Verified, +1.9% on Multi-SWE-Bench, and 37% faster task completion. Emergent capabilities include the 'architect mindset' (pre-planning code structure) and a Forge-based RL training framework using 200,000+ environments. M2.7 (April 2026) builds on M2.5 by adding interleaved reasoning chains and multi-tool orchestration. M2.5 remains the most documented and battle-tested release.

re-authored by TheRouter
What is the Forge RL framework?

Forge is MiniMax's proprietary agent-native reinforcement learning framework. It introduces an intermediary layer that decouples the training-inference engine from the agent scaffold, supporting arbitrary agent integration. Key innovations: asynchronous scheduling for throughput vs off-policyness balance, tree-structured merging achieving ~40x training speedup, and CISPO (Clipping Importance Sampling Policy Optimization) algorithm for stable MoE training during large-scale RL.

re-authored by TheRouterwww.minimax.io β†—
How much does MiniMax M2.5 cost?

Through TheRouter, M2.5 costs $0.40 per million input tokens and $1.60 per million output tokens. The Highspeed variant (minimax/m2.5-highspeed, ~100 TPS) costs $0.80/$3.20. MiniMax's own API pricing is $0.15/$1.15 (standard) and $0.30/$2.40 (Lightning). Continuous operation cost: ~$0.30/hour at 50 TPS or ~$1.00/hour at 100 TPS.

re-authored by TheRouter
Can I use MiniMax M2.5 commercially?

Yes. M2.5 is released under a Modified MIT License. Commercial use is permitted with the requirement to 'prominently display MiniMax M2.5 on the user interface of such product or service'. This is the same license used by M2.1. Self-hosting is allowed.

re-authored by TheRoutergithub.com β†—
How do I use M2.5 via TheRouter?

M2.5 is available at the model ID minimax/m2.5 through TheRouter's standard OpenAI-compatible endpoint at https://api.therouter.ai/v1/chat/completions. It supports all standard parameters: temperature (0-2), max_tokens, top_p, tools, tool_choice, response_format (json_object/json_schema), stop, and streaming. For higher throughput, use minimax/m2.5-highspeed (same API shape, ~100 TPS at 2x cost).

re-authored by TheRouter
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release datewww.minimax.io β†—2026-05-27verified
Architecturewww.minimax.io β†—2026-05-27verified
Licensegithub.com β†—2026-05-27verified
RL frameworkwww.minimax.io β†—2026-05-27verified
Training environmentswww.minimax.io β†—2026-05-27verified
Training languageswww.minimax.io β†—2026-05-27verified
SWE-bench Verifiedwww.minimax.io β†—2026-05-27verified
Multi-SWE-benchwww.minimax.io β†—2026-05-27verified
BrowseCompwww.minimax.io β†—2026-05-27verified
BFCL (Tool Calling)www.minimax.io β†—2026-05-27verified
VIBE-Prowww.minimax.io β†—2026-05-27to verify
SWE-bench Multilingualwww.minimax.io β†—2026-05-27verified
Wide Searchwww.minimax.io β†—2026-05-27to verify
RISEwww.minimax.io β†—2026-05-27to verify
GDPval-MMwww.minimax.io β†—2026-05-27to verify
MiniMax releases M2.5 -- a frontier model matching Claude Opus 4.6 at 1/10th the costwww.minimax.io β†—2026-05-27verified
What is the Forge RL framework?www.minimax.io β†—2026-05-27to verify
Can I use MiniMax M2.5 commercially?github.com β†—2026-05-27to verify
Help & contact